JPH0632011B2 - Voice recognizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention will be described in the following order.

A Industrial field of use B Outline of the invention C Conventional technology D Problems to be solved by the invention E Means for solving problems (Fig. 1) F Action G Example G1 Description of acoustic analysis circuit (No. 1) 2) Explanation of G2 time normalization processing (FIGS. 2 and 3) Explanation of G3 pattern matching processing (FIG. 2) Explanation of G4 digital filter (8) (FIGS. 4-9)
Fig. H Effect of the invention A Industrial field of application The present invention is a device for performing voice recognition by performing pattern matching between a standard pattern of a recognition target word that has been created and stored in advance and an input pattern of a word to be recognized. Regarding

B Outline of the Invention In the present invention, in a device that performs voice recognition by pattern matching, as a pattern for matching, the trajectory of an acoustic parameter time series obtained by acoustic analysis in a voice section of an input voice signal is estimated in the parameter space. Then, the new recognition parameters obtained by re-sampling the trajectory at a predetermined interval are used, and the acoustic parameter time series signals are passed through a low-pass filter so that the recognition parameter time series can be obtained even if there is fluctuation in the stationary part of the input speech. Has almost no influence.

C Conventional technology Speech is a phenomenon that changes along the time axis,
A unique word or words are created by uttering a voice so that the pattern changes every moment. Speech recognition is the technology that automatically recognizes words and words spoken by humans.
At present, it is extremely difficult to realize voice recognition that is comparable to human auditory function. Therefore, most of the speech recognition currently in practical use is a method of performing pattern matching between a standard pattern of a recognition target word and an input pattern under a certain use condition.

FIG. 10 is a diagram for explaining the outline of the voice recognition device, and voice input from the microphone (1) is supplied to the acoustic analysis circuit (2). In this acoustic analysis circuit (2), acoustic parameters representing the characteristics of the input voice pattern are extracted. Various methods of acoustic analysis for extracting the acoustic parameters are conceivable. For example, as an example, a bandpass filter and a rectifier circuit are provided as one channel, and a plurality of such channels are arranged with different passbands. A method of extracting a time change of a spectrum pattern as an output of a group is known. In this case, the acoustic parameters are the time series Pi (n) (i = 1, 2 ... I; I
Is the number of channels of the bandpass filter, n = 1,
2 ... N; N can be represented by the number of frames used for recognition in the section determined by the voice section determination.

The acoustic parameter time series P from this acoustic analysis circuit (2)
i (n) is supplied to a mode switching circuit (3) including a switch, for example. When the switch of this circuit (3) is switched to the terminal A side, the acoustic parameter time series Pi (n) is stored in the standard pattern memory (4) as a recognition parameter in the registration mode. That is, the voice pattern of the speaker is stored in this memory (4) as a standard pattern prior to the voice recognition. At the time of registration, generally, the number of frames of each registered standard pattern is different due to the variation in vocalization speed and the difference in word length.

On the other hand, when the switch (3) is switched to the terminal B side, it is in the recognition mode. Then, in this recognition mode, the acoustic parameter time series of the input voice at that time from the acoustic analysis circuit (2) is supplied to the input voice pattern memory (5) and temporarily stored. Then, the magnitude of the difference between this input pattern and each of the standard patterns of the plurality of recognition target words read from the standard pattern memory (4) is calculated by the distance calculation circuit (6). The recognition target word having the smallest difference from is detected by the minimum value judgment circuit (7), and the input word is recognized.

In this way, the input voice is recognized by the pattern matching process between the registered standard pattern and the input pattern. In this case, even if the same word is uttered in the same way, the spectrum pattern shifts in the time axis direction. It must be taken into consideration that it expands and contracts. That is, for example, when recognizing the word "high", when the standard pattern is registered as "high", and the input voice extends "hi" in the time axis direction, this is a big difference in distance. It is a completely different word, and I cannot recognize it correctly. Therefore, in pattern matching for voice recognition, it is necessary to perform time normalization processing for correcting the displacement and expansion / contraction in the time axis direction, and this time normalization is an important processing for improving recognition accuracy. is there.

As one method of this time normalization, DP (Dynamic)
There is a method called programming matching (see, for example, Japanese Patent Laid-Open No. 50-96104).

This DP matching does not prepare a large number of standard patterns in consideration of the shift of the time axis, but generates a standard pattern in which a large number of times are normalized by a distortion function and calculates the distance between this and the input pattern. The voice recognition is performed by detecting the minimum value.

By the way, when this DP matching method is used, the number of frames of standard patterns to be registered is indefinite, and it is necessary to perform DP matching processing between all registered standard patterns and input patterns. It has the drawback of increasing dramatically.

Further, since the DP matching is a matching method that emphasizes the stationary part (the part where the spectrum pattern does not change with time), there is a possibility that erroneous recognition may occur between the partially similar patterns.

The present applicant has previously proposed a method of time normalization that does not cause such a defect (for example, Japanese Patent Application No. 59-106177).
issue).

That is, the acoustic parameter time series Pi (n) draws a sequence of points when the parameter space is considered. For example, when the recognition target word is “HAI” and the number of bandpass filters for acoustic analysis is two, and Pi (n) = (P ₁ P ₂ ), the acoustic parameter time series of the input speech is the two-dimensional A series of points as shown in FIG. 11 is drawn in the parameter space. As is apparent from this figure, the point sequence of the non-stationary part of the voice is roughly distributed, and the quasi-stationary part is densely distributed. In this case, if the voice is completely stationary, the parameters will not change, and in that case the point sequence will stop at one point in the parameter space, but even if a human produces the same sound, Therefore, it does not become a complete steady state, and the influence of fluctuation appears as a quasi-stationary part as shown in the figure.

From the above, it is considered that most of the deviation in the time axis direction due to the fluctuation of the vocalization rate of the voice is due to the difference in the point sequence density of the quasi-stationary part, and the influence of the time length of the non-stationary part is small. Therefore, if a locus drawn by a continuous curve that approximately passes through the entire point sequence as shown in FIG. 12 is estimated from the point sequence of the input parameter time series Pi (n), this locus will produce a voice. It can be seen that it is almost invariant to speed fluctuations.

Therefore, the applicant has proposed the following time axis normalization method. That is, first, the input parameter time series P
A continuous curve from the start Pi (l) to the end Pi (N) of i (n) Estimate the trajectory drawn in. In this case, the trajectory is estimated by, for example, linearly approximating the acoustic parameter time series as shown in FIG. This estimated curve The length S of the locus is obtained from Then, as shown by the mark ◯ in FIG. 13, re-sampling is performed at a predetermined length T along this locus. For example, when re-sampling to M points, the locus is re-sampled based on the length of T = S / (M-1) ... (1). Let Qi be the parameter time series that draws this resampled point sequence.
(M) (i = 2, 2 ... I, m = 1, 2 ... M), this parameter time series Qi (m) has basic information of the locus, and It is a parameter that is almost invariant to changes in vocalization rate. That is, it is a recognition parameter time series whose time axis is normalized.

Therefore, this parameter time series Qi (m) is registered as a standard pattern, and an input pattern is also obtained as this parameter time series Qi (m), and the distance between both patterns is obtained by this parameter time series Qi (m). ,
If the voice with the smallest distance is detected and the voice is recognized, the voice is always recognized with the deviation in the time axis direction being normalized and removed.

In addition, according to this processing method, the recognition parameter time series Qi
Since the number of frames in (m) is always M, and the recognition parameter time series Qi (m) is time-normalized,
Even if the calculation of the distance between the input pattern and the registered standard pattern is the simplest Chebyshev distance calculation, a good effect can be expected.

Further, the above method is a method of time normalization in which the non-stationary part of the voice is more emphasized, and erroneous recognition between partially similar patterns such as DP matching processing is reduced.

Furthermore, the variation information of the vocalization rate is not included in the normalized parameter time series Qi (m), which facilitates the handling of the global characteristics of the parameter transition structure coordinated in the parameter space, and enables the unspecified speaker recognition. It is possible to apply various effective methods.

In the following, the time normalization processing described above is performed by the NAT.
(Normalization Along Traj
This is called processing.

D Problem to be Solved by the Invention As described above, even if the input voice is a monotone, the acoustic parameter Pi (n) from the acoustic analysis circuit does not reach a steady state and fluctuates as shown in FIG. Therefore, NAT
It is affected by this fluctuation when reconstructing a trajectory in processing. For example, when linear interpolation is performed as shown in FIG. 13, the magnitude of this fluctuation directly affects the trajectory length. Therefore, the normalized recognition parameter time series Qi (n) has an error due to this fluctuation, which leads to a decrease in the voice recognition rate.

E Means for Solving the Problems FIG. 1 is a diagram showing an example of the basic configuration of the speech recognition apparatus according to the present invention, and the portions corresponding to those in FIG.

In this example, the acoustic analysis circuit (2) uses a bandpass filter group. That is, the audio signal from the microphone (1) is supplied to the A / D converter (21) and converted into a digital signal, and this digital signal is supplied to the digital bandpass filter group (22) and is composed of a plurality of frequency components. The signal is converted. The output of this band pass filter group (22) is the feature extraction circuit (23).
Is supplied to. The digital voice signal from the A / D converter (21) is also supplied to the voice section determination circuit (25) to determine the section in which the voice is input to the microphone (1), and the determination output is the determination output. (23).

In the feature extraction circuit (23), the acoustic parameter time series Pi (n) is created from the output of the bandpass filter group (22) in this voice determination section, and this is used as the output of the acoustic analysis circuit (2).

Acoustic parameter time series Pi in this voice determination section
(N) is supplied to a digital filter (8) having a low pass filter characteristic. The filter characteristic of the digital filter (8) may be such that it can pass up to 0.3 times the output band of the bandpass filter group of the acoustic analysis circuit (2) in principle. The output of this digital filter (8) is the acoustic parameter time series P in which the fluctuation of the stationary part is suppressed.
i (n) 'is obtained and supplied to the NAT processing circuit (9).

In the NAT processing circuit (9), the trajectory in the acoustic parameter space is estimated from the acoustic parameter time series Pi (n) 'as described above, and a new recognition parameter time series Qi (m) is formed based on this trajectory. To be done.

The parameter time series Qi (m) is stored and registered in the standard pattern memory (4) in the registration mode through the mode switching circuit (3) and is supplied to the distance calculation circuit (6) in the recognition mode. , The distances from the standard pattern memory (4) to a plurality of registered standard patterns are calculated, the minimum standard pattern of the calculation result is judged by the minimum value judgment circuit (7), and the judgment output is made the recognition output. It

Actually, the NAT process is performed by using a microcomputer. In this case, when estimating the trajectory from the acoustic parameter time series Pi (n) ′ in the voice determination section,
Instead of using the starting point Pi (1) 'of the parameter time series Pi (n)' as the starting point of the locus, it is also possible to estimate the silent point as the starting point without fail as in the example of the figure. The same can be applied to the terminal Pi (N).

F action acoustic parameter time series Pi from acoustic analysis circuit (2)
(N) passes through a digital filter (8) having a low-pass filter characteristic, and the fluctuation of the steady part is reduced. Since this is supplied to the NAT processing circuit (9) and used for trajectory estimation, the influence of fluctuations on the trajectory is suppressed, and it is expected that the recognition rate of voice will be improved as a new recognition parameter time series Qi (n). Is obtained.

G. Embodiment FIG. 2 shows an embodiment of the speech recognition apparatus according to the present invention, in which a 16-channel bandpass filter group is used for acoustic analysis.

G1 Description of Acoustic Analysis Circuit (2) That is, in the acoustic analysis circuit (2), the audio signal from the microphone (1) is an A / D converter via an amplifier (211) and a low-pass filter (212) for band limitation. It is supplied to (213) and converted into a 12-bit digital audio signal at a sampling frequency of 12.5 kHz, for example. This digital audio signal is converted into digital bandpass filters (221 ₁ ) and (22 ₁ ) of each channel of a 15-channel bandpass filter bank (22).
1 _2), ..., it is supplied to the _{(221 16).} This digital bandpass filter (221 ₁ ), (22
1 _2), ..., _{(221 16)} for example is constituted by Butterworth fourth order digital filter, the band band from 250Hz to 5.5KHz is divided at equal intervals on a logarithmic axis each filter It is designed to be the pass band of. Then, each digital bandpass filter (221
₁ ), (221 ₂ ), ..., (221 ₁₆ ) output signals are rectification circuits (221 ₁ ), (222 ₂ ),
..., (221 ₁₆ ) are supplied to these rectifier circuits (222 ₁ ), (222 ₂ ), ..., (222
₁₆ ) outputs the digital low-pass filter (2
23 ₁ ), (223 ₂ ), ..., (223 ₁₆ ). These digital low-pass filters (223
₁ ), (223 ₂ ), ..., (223 ₁₆ ) are configured by, for example, FIR low-pass filters having a cutoff frequency of 52.8 Hz.

Each digital low-pass filter (223 ₁ ), (223 ₂ ), ..., (22) which is the output of the acoustic analysis circuit (2)
The output signal of 3 ₁₆ ) is supplied to the sampler (231) which comprises the feature extraction circuit (23). This sampler (231) has a digital low-pass filter (22
The output signals of 3 ₁ ), (223 ₂ ), ..., (223 ₁₆ ) are sampled at every frame period of 5.12 mesc. Therefore, rather than this, the sample time series Ai
(N) (i = 1, 2, ..., 16; n is a frame number, n = 1, 2, ..., N).

The output from the sampler (231), that is, the sample time series Ai (n) is supplied to the sound source information normalization circuit (232), which eliminates the difference in vocal cord sound source characteristics depending on the speaker of the voice to be recognized. It

That is, the logarithmic transformation of i (n) = log (Ai (n) + B) (2) is performed on the sampler time series Ai (n) supplied from the sampler (231) for each frame period. In this equation (1), B is set to a value such that the noise level is hidden by the bias.

Then, when the vocal cord sound source characteristic is approximated by the expression yi = a · i + b, the coefficients a and b are determined by the following expressions.

When the normalized parameter of the sound source is Pi (n), the parameter Pi (n) is Pi (n) = i (n)-{a (n) .i + b (n) when a (n) <0. )}
It is expressed as (5).

When a (n) ≧ 0, only the level normalization is performed, and the parameter Pi (n) is Is expressed as

The sound parameter information normalization circuit (232) obtains the acoustic parameter time series Pi (n) in which the difference in vocal cord sound source characteristics is normalized and removed.

This acoustic parameter time series Pi (n) is supplied to the digital filter (8). The digital filter (8) is an interpolation filter having a low-pass filter characteristic as described later, and fluctuations in the steady part of the locus estimated by the low-pass filter characteristic of the interpolation filter (8) in the NAT processing circuit (9) in the subsequent stage. Is removed, and the acoustic parameter Pi (n) is interpolated so that more accurate trajectory estimation can be performed. The interpolation is performed by inserting (P-1) pieces of data of "0" between each piece of the acoustic parameter Pi (n) and then performing no FIR filtering, whereby the acoustic parameter Pi in which the number of data is increased P times. (N) 'is obtained from this.

In this way, the number of data samples is increased by a factor of P, and the acoustic parameter Pi (n) ′ from which fluctuations in the stationary part have been removed by the low-pass filter characteristic is supplied to the intra-speech interval parameter memory (200). The parameter memory (200) in the voice section receives the voice section determination signal from the voice section determination circuit (24) and receives the parameter Pi (n).
Is stored for each determined voice section.

The voice section determination circuit (24) is a zero cross counter (24
1), a power calculation circuit (242) and a voice section determination circuit (243), and a digital voice signal is supplied from an A / D converter (213) to a zero cross counter (241) and a power calculation circuit (242). In the zero cross counter (241), one frame period is 5.12 mse.
For each c, the number of zero-crosses of the digital audio signal of 64 samples within this one frame period is counted, and the count value is supplied to the first input terminal of the audio section determination circuit (243). The power calculation circuit (242) obtains the power of the digital voice signal within this one frame period, that is, the sum of squares, for each frame period, and the output power signal is the second input end of the voice section determination circuit (243). Is supplied to. The sound source normalization information from the sound source information normalization circuit (232) is further supplied to the third input terminal of the voice section determination circuit (243). Then, in the voice section determination circuit (243), the number of zero crosses, the intra-section power, and the sound source normalization information are processed in a complex manner, and the process of determining silence, unvoiced sound, and voiced sound is performed to determine the voice section.

The voice section determination circuit (243) outputs the voice section determination signal indicating the determined voice section to the voice section determination circuit (2).
4) as an output of the voice section parameter memory (20
0).

In this way, the acoustic parameter time series Pi (n) ′ stored in the memory (200) in the judgment voice section is NA.
It is supplied to the T processing circuit (9).

Explanation of G2 time normalization processing The NAT processing circuit (9) comprises a trajectory length calculation circuit (91), an interpolation interval calculation circuit (92) and an interpolation point extraction circuit (93).

Parameter time series P from the parameter memory (200)
i (n) '(i = 1, 2, ..., 16; n = 1,
2, ..., N) are supplied to the trajectory length calculation circuit (91). In the locus length calculation circuit (91), the length of the locus by the linear approximation drawn by the acoustic parameter time series Pi (n) 'in the parameter space as shown in FIG. 13 is calculated.

In this case, the Euclidean distance D (a _i , b _i ) between the I-dimensional vectors a _i and b _i is Is. Therefore, the I-dimensional acoustic parameter time series Pi
From (n) ', the distance S (n) between the parameters adjacent to each other in the time series direction when the trajectory is estimated by the linear approximation is S (n) = D (Pi (n + 1)', Pi (n) ') (n = 1, ..., N) ... (8) Then, from the first parameter Pi (1) 'to the nth parameter Pi in the time series direction.
The distance SL (n) to (n) 'is Is represented. Note that SL (1) = 0.

And the total locus length SL is Is represented. The trajectory length calculation circuit (91) uses this (11)
The signal processing shown in equations (12) and (13) is performed.

The locus length SL obtained by the locus length calculation circuit (91)
Is supplied to the interpolation interval calculation circuit (92).
This interpolation interval calculation circuit (92) calculates the resampling interval T when resampling along the locus.

In this case, if the resampling is performed at the point M, the resampling interval T is obtained as T = SL / (M-1) ... (11).

A signal indicating the resampling interval T from the interpolation interval calculation circuit (92) is supplied to the interpolation point extraction circuit (93). The acoustic parameter time series Pi (n) 'from the parameter memory (200) is also supplied to the interpolation point extraction circuit (93). This interpolation point extraction circuit (93)
Is resampled at the resampling interval T along the locus of the acoustic parameter time series Pi (n) 'in that parameter space, for example, the locus obtained by linear approximation between parameters, as shown by the circles in FIG. From the new sequence of points obtained by
(M) is formed.

Here, in the interpolation point extraction circuit (93), processing according to the flowchart shown in FIG. 3 is performed, and a recognition parameter time series Qi (m) is formed.

First, in step [101], the value 1 is set to the variable J indicating the number of the resampling points in the time series direction, and the value 1 is set to the variable IC indicating the frame number of the acoustic parameter time series Pi (n) '. And is initialized. Next, in step [102], the variable J is incremented, and in step [103], it is determined whether or not the variable J at that time is (M-1) or less. Determine if the number in the time series direction is the last number that needs to be resampled. If it is the last number, the process proceeds to step [104], and the resampling ends.

If it is not the last number, the first re-sampling point (this is always a silent portion) in step [105].
To the J-th resampling point are calculated. Next, in step [106], the variable IC is incremented. Next step [10
7], the re-sampling distance DL is the acoustic parameter time series P.
Distance SL from the first parameter Pi (l) 'of i (n)' to the _ICth parameter Pi _{(lC ')
It} is determined whether or not the resampling point at that time is located closer to the starting point side of the locus than the parameter Pi _{(lC ') at} that time depending on whether it is smaller than _(lC'). If not, step [1
06], the variable IC is incremented, and then the re-sampling point and the parameter Pi are again determined in step [107].
The position on the locus with _{(lC ′)} is compared, and the resampling point is the parameter Pi _{(lC ′)} on the locus.
When it is determined that the recognition parameter Qi _(J) is located closer to the start point side than that, the process proceeds to step [108] to form the recognition parameter Qi _(J) .

That is, the (IC-1) th parameter Pi located closer to the start point than the Jth resampling point is from the resampling distance DL at the Jth resampling point.
_The distance SL _(lC-1) by _(lC-1) 'is subtracted to obtain the distance SS from the (IC-1) th parameter Pi _(lC-1) ' to the _Jth resampling point.
Next, the distance S (n) between the parameter Pi _(lC-1) 'and the parameter Pi _(lC') located on both sides of this J-th resampling point on the locus (this distance S
(N) is obtained by the signal processing represented by the equation (11). ), Divide this distance SS, and divide by SS / S
_The parameters Pi _{(lC ′)} and Pi located on both sides of the _Jth resampling point on the locus at _(IC-1).
Difference from _(lC-1) '(Pi _(lC')- Pi
_(IC-1) ') is multiplied to obtain the (IC-1) th parameter Pi _(lC- ) located adjacent to the start point side of the _Jth resampling point on the _{locus. 1)} ′ is calculated, and the interpolation amount and the (IC−1) th parameter Pi located adjacent to the start point side of the Jth resampling point and the interpolation amount.
_(LC-1) 'is added to form a new recognition parameter Qi _(J) along the trajectory.

In this way the start and end points (when they are each silent Is. ) Is resampled to form a recognition parameter time series Qi (m).

G3 Pattern Matching Process Description The recognition parameter time series Qi (m) from the NAT processing circuit (9) is stored in the standard pattern memory (4) for each recognition target word in the registration mode by the mode changeover switch (3). It Further, in the recognition mode, the distance is supplied to the distance calculation circuit (6) and the distance between the standard pattern memory (4) and the parameter time series of the standard pattern is calculated. The distance in this case is calculated as a simple Chebyshev distance, for example. The calculation output of the distance between each standard pattern and the input pattern from the distance calculation circuit (6) is supplied to the minimum value determination circuit (7), the standard pattern having the minimum distance calculation value is determined, and the determination result is determined by this determination result. The recognition result of the input voice is obtained at the output end (70).

Description of G4 Interpolation Filter (8) FIG. 4 shows an example of the configuration of the interpolation filter (8). The parameters P ₁ (n), P of each channel are shown.
_A zero data insertion circuit (8) for packing 0 data between data for each of ₂ (n), ..., P ₁₆ (n)
1 _{1) (81} 2) ... _{(81 16)} and the FIR filter _{(82 1)} (82 2) ... _{(82 16)} are provided.

In the zero data insertion circuits (81 ₁ ) to (81 ₁₆ ), as shown in FIG. 5A, between the adjacent parameters Pi (k) and Pi (k + 1), there are P-1 pieces of 0 data indicated by a circle. For example, three pieces are inserted.

Therefore, from the zero data insertion circuit (81i), Pi (n) '= [Pi (0), φ, φ, φ, Pi
(1), φ, φ, φ, Pi (2) ...] The parameter time series Pi (n) ′ is obtained. That is, the parameter time series Pi (n) 'has the number of sample data increased by P = 4 times that of the parameter time series Pi (n).

This new parameter time series Pi (n) 'is supplied to each of the FIR filters (82 ₁ ) to (82 ₁₆ ).

Each of the FIR filters (82 ₁ ) to (82 ₁₆ ) is configured as shown in FIG. 6, for example.

That is, in the figure, (820) is an input terminal, and (821 ₁ ) (821 ₂ ) ... (821 _J ) are delay elements for each unit time. In this example, the unit time is the frame period. It is set to 1/4 (= 1 / p). Also,
(822 ₀ ) (822 ₁ ) (822 ₂ ) ... (82
2 _J ) is a multiplier, and the data and delay elements (821 ₁ ) (821 ₂ ) ...
The data obtained at the output of (821 _J ) is multiplied by the filter coefficient αj (j = 0, 1, 2, ..., J). Then, the outputs of these multipliers (822 ₀ ) to (822 _J ) are supplied to the adder circuit (823), and the added output is obtained at the output terminal (824). Therefore, at the output terminal (824), the parameter data Pi (k) and Pi
Between (k + 1) and the input parameter P, as indicated by the ● mark in FIG.
An output is obtained from i (n) in which three pieces of sample data are interpolated. That is, the parameter time series P ₁ (n) ′, P with the number of parameters increased four times
₂ (n) ... P ₁₆ (n) 'are respectively obtained and stored in the parameter memory (200), and the NAT processing is performed based on this.

Then, FIR filter ₍₈₂ 1) of this case - (82
₁₆ ) each of the multipliers (822 ₀ ) to (822 _J )
Each of the FIR filters (82 ₁ ) to (82 ₁₆ ) has a low-pass filter characteristic by selecting the multiplication coefficient α _j of. Moreover, α ₀ =
_{α J, α 1 = α J} -1, is selected as _{α 2 = α J-2 ····} , the phase characteristic is linear, the group delay characteristic is to be constant for all frequencies.

The low-pass filter characteristic at this time is as shown in FIG. 7, and the cutoff frequency is π / 2P × β (0 <β ≦ 1).
Therefore, when β = 1, there is no low-pass filter characteristic, and as β approaches 0, the low-pass filter characteristic becomes steeper. That is, the pass band of the low pass filter is [0, Ω / 2 × β]. Ω is the sampling frequency of the original parameter time series Pi (n).

As described above, the interpolation filter (8) processes the parameter time series Pi (n) to increase the number of data and pass the low pass filter to eliminate the fluctuation in the stationary part.

Therefore, this can be expected to improve the voice recognition rate.

For example, when I = 16, P = 4, the order of the low-pass filter is 129, and β = 0.4 (shown by the solid line in the figure)
FIG. 8 shows the difference in recognition rate between the case where β = 1 (low-pass filter through) (shown by the broken line in the figure). Also, the acoustic parameter time series Pi (n) of the first channel
FIG. 9A and FIG. 9B show the output changes when the low pass filter is not passed and when it is passed.

As is clear from FIG. 8, the number of registered people is 9 and the number of vocalizations (the number of registrations) is 3 times as compared with the case where only the NAT processing is performed.
The highest recognition rate was 97.45% against 96.24%, which was an improvement of 1.21%.

It should be noted that this is not only because the effect of trajectory fluctuation is eliminated by the low-pass filter characteristic, but also by interpolation by NAT.
This is also due to the fact that the number of input data of the processing circuit (9) is large and the error due to interpolation in the NAT processing is small.

H Effect of the Invention As described above, according to the present invention, the low-pass filter is provided before performing the NAT processing, and the high frequency component of the acoustic parameter output from the acoustic analysis circuit is cut, so that the acoustic parameter output becomes steady. Fluctuations in parts are eliminated. This reduces the error in the trajectory estimation in the NAT processing circuit, reduces the error in the recognition parameter time series to be created, and improves the voice recognition rate.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the device of the present invention, FIG. 2 is a block diagram of a specific embodiment of the device of the present invention, and FIG. 3 is a flow chart for explaining the operation of the main part thereof. , FIG. 4 and FIG. 6 are block diagrams showing the configuration of an example of the main circuit, FIG. 5 and FIG. 7 to FIG. 9 are diagrams for explaining the same, and FIG. A block diagram showing the configuration, and FIGS. 11 to 13 are diagrams for explaining the NAT processing. (2) is an acoustic analysis circuit, (4) is a standard pattern memory,
(6) is a distance calculation circuit between the standard pattern and the input pattern, (7) is a minimum value determination circuit, (8) is a digital filter, and (9) is a NAT processing circuit.

Front page continuation (72) Inventor Makoto Akabane 6-735 Kitashinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor Masao Watanabe 6-735 Kitashinagawa, Shinagawa-ku, Tokyo Sony Corporation Shares In-house (56) References Acoustical Society of Japan Proceedings Proceedings October 1984 1-9-9 p. 17-18

Claims (1)

[Claims] 1. A speech analysis means for obtaining a time series of acoustic parameters of an input speech signal, a trajectory drawn by the acoustic parameter time series from the acoustic analysis means in a parameter space is estimated, and resampling is performed along the trajectory. A time normalization means for obtaining a time-normalized recognition parameter time series, a standard pattern memory in which a recognition parameter time series of a standard pattern of a recognition target word is stored, and a recognition parameter of an input pattern from the time normalization means Distance calculating means for calculating the difference between the time series and the recognition parameter time series of the standard pattern from the standard pattern memory, and the minimum value for detecting the minimum value calculated by the distance calculating means to obtain a recognition output. Determining means, and the time normalization means for the acoustic parameter time series from the acoustic analysis means through a low-pass filter. A speech recognition apparatus adapted to remove the influence of fluctuation in the quasi-stationary part of the locus by supplying